First hands-on results of Google Gemini 4 Carbon surface: 3D generation stronger than Argon but 2 to 4 times slower

It's only been 8 days since Argon fully opened up, and Google's next ace has already charged into the arena!

Just moments ago, Google's strongest model yet, Carbon, was revealed to be in testing. Partners have already gotten their hands on it first, blind tests have opened on LMArena, and release is right around the corner.

As soon as the news broke, testers all rushed into the Arena.

Within a single day, Carbon hands-on tests exploded across X, with some people using it to complete an entire level of Super Mario in one go.

The previous generation flagship's hype hasn't even cooled off, and the next generation is already lining up at the door.

Google co-founder Brin said Google has begun stepping onto the "RSI treadmill," with models helping to build the next generation of models.

Super Mario cleared in one run, Carbon's 3D is absurdly strong

Carbon on the Arena still carries the name "gemini-4-argon." The way testers recognize it is rather amusing—they just watch whether it's slow.

Netizen Harshith discovered that there are actually two models under this name. One is fast and decent, the other is unusually slow but its output is truly stunning—the latter is Carbon.

Using a Super Mario prompt, the review account Lumina had Carbon build an entire level of the game.

The protagonist goes along bonking bricks, collecting coins, and entering pipes, descends into the underground level and climbs back out, finally jumps onto the flagpole and walks into the castle, with "COURSE CLEAR" popping up on screen and the coin count and rating settled out.

Lumina also put the same voxel pagoda prompt to Argon, Carbon, Opus 5.5, and Sonnet 5.5 for a head-to-head comparison.

In Lumina's view, Carbon's quality is clearly higher than Argon's, but depending on the task, it's 2 to 4 times slower.

Netizen Synopsis's voxel valley pushed the visuals up another level.

Cherry blossoms bloom along the riverbank, a red arched bridge spans the water, wooden houses have their lights on, a rain shower starts out of nowhere, and the whole valley darkens along with it.

Netizen Salio built a 3D castle whose lighting can switch between dusk, blue hour, moonlight, and misty morning light, with the stone bridge below lit all the way along.

Well-known leaker chetaslua handed Gemini's mascot to Carbon and had it make a 3D animation explaining why carbon is stable.

At the start, the mascot stares at a carbon-12 nucleus. Then 6 electrons arrange as 1s²2s²2p², the orbitals hybridize into 4 identical sp³ lobes, which then share electrons with neighbors to reach 8, finally building up an entire diamond lattice.

chetaslua said this prompt was extremely difficult for previous Gemini models. Using Carbon to explain carbon—Google's naming this time has been thoroughly figured out by netizens.

Netizen Salio built a 3D castle whose lighting can switch between dusk, blue hour, moonlight, and misty morning light, with the stone bridge below lit all the way along.

Developer Pankaj Kumar had it make an interactive 3D crystal citrus.

However, when it comes to SVG, evaluations start to diverge.

Netizen Binkg had Carbon draw an SVG of a cola bottle, the body covered in water droplets, with the small text on the label and the bottle cap all drawn out, and frost, bubbles, and backlight adjustable on the side.

But after watching tests all day, Haleemah said Carbon's strongest area is still 3D, and she hasn't yet seen a good SVG.

Beyond SVG, the frontend is still the same old problem. Lumina said Carbon's frontend design and aesthetics remain mediocre, lacking that flavor of Anthropic's models.

The same model can surge in 3D yet stand still on the frontend because the two types of tasks are tested differently.

In 3D scenes, geometry, lighting, and collision are all determined by code—get one thing wrong and the image collapses, and whether it's right can be verified at a glance.

The frontend tests how much whitespace, what font size, where buttons go—there's no standard answer, only human preference.

So 3D can be pushed upward all the way through compute and reinforcement learning, while the frontend can only be slowly honed through massive human feedback.

Slower but not lazy, Carbon specifically cures Argon's laziness

When testers show off their work, almost all of them add a note that it's slow, but few complain, because it's no longer lazy.

Laziness, as it happens, is the flaw Argon has been most harshly criticized for since launch. Its benchmark scores are beautiful, but within less than a week of developers getting hands-on, its reputation for "laziness" spread.

After testing Argon, tech media outlet Every gave 5 usage tips, and the 3rd was to make sure to double-check its work.

In 38 benchmark runs, 17 times Argon turned the code review step on and then off again, with only 30 milliseconds in between.

Tucked into the same tip was another set of numbers. Raising Argon from a low reasoning tier to a high one only made it take 1.7 times as long; raising Opus 5.5 to a high tier takes 18 times as long.

In other words, telling Argon to think more doesn't make it think much more.

On Every CEO Dan Shipper's senior engineer benchmark, Opus 5.5 and Sonnet 5.5 both scored 85%, GPT-6 Astra scored 81%, and Argon scored only 21%, dead last among 8 models.

But with Carbon, the situation flips completely.

Business Insider previously reported that after trying Carbon internally, Google employees said writing code with it "feels a lot like Opus 5.5."

Carbon emerges right after Argon rolls out, Google steps onto the RSI treadmill

From Argon to Carbon, Google's pace this round is astonishingly fast.

Argon was officially announced on September 30, initially given only to cyber defense parties in the Fairwind program; on October 13, Argon fully opened to developers and subscribers; just 8 days later, Carbon entered partner testing.

Google is playing its cards faster and faster, and the answer Brin gave is—RSI (recursive self-improvement).

In a recent roundtable discussion, he said that researchers are now already hand in hand with models, building the next version of the model together.

Argon's safety report shows that in September of this year, in over 90% of DeepMind's code merge requests, the AI-written portion was above 75%.

Merge requests fully handed to AI, with humans interjecting fewer than 3 rounds, also rose from 7.5% in April to 11.9%.

DeepMind's monthly code merge requests—the darker the color, the higher the proportion written by AI

When models start helping to build models, it's not hard to understand why Argon and Carbon are so close together.

Source of this article: AI Era Risk Warning and Disclaimer

Markets carry risk, and investment requires caution. This article does not constitute personal investment advice, nor does it take into account the specific investment objectives, financial situation, or needs of individual users. Users should consider whether any opinions, views, or conclusions in this article fit their particular circumstances. Invest at your own risk based on this.